A lightweight Efficient-UNet model was developed by integrating the U-Net model of convolutional neural network with the Efficient-Net framework. Taking the Liulin Experimental Watershed as a case study, radar echo extrapolation with 1h lead time was performed for representative rainfall events. Based on measured precipitation data, quantitative precipitation forecasts were further conducted using the global dynamic correction, dynamic intensity-class correction, and dynamic regional correction methods. The forecast results of the Efficient-UNet model were systematically compared with those of the U-Net model, the SmaAt-UNet model, and the Simple-UNet model with a three-layer encoder-decoder structure. The results showed that, within 1h lead time, the Efficient-UNet model achieved radar echo extrapolation performance comparable to the U-Net model, and exhibited superior capability in predicting the spatial extent and temporal evolution of radar echoes. The SmaAt-UNet model provided higher accuracy for strong echo, whereas the Simple-UNet model yielded the poorest overall extrapolation performance. The precipitation forecasting accuracies obtained using the dynamic intensity-class and dynamic regional correction methods were comparable, and both outperformed the global dynamic correction method. The precipitation forecasting accuracy of the Efficient-UNet model was generally consistent with the U-Net model and superior to those of the other models, whereas, the Efficient-UNet model required substantially fewer model parameters and lower computational resources. Keywords: precipitation forecasting; convolutional neural network; weather radar; echo extrapolation; U-Net model；Efficient-Net model; the Liulin Experimental Watershed 〖FL